Resume screening

When to use this scenario

Resume screening produces a structured fit score plus a summary of a candidate's qualifications relative to a job description. The goal is not to make hiring decisions but to help recruiters triage 500 applications down to 30 worth reading carefully. Speed and throughput matter: a company posting a senior engineer role may receive 1,000 applications in 48 hours.

Gemini 2.5 Flash handles this volume economically — at $0.075/million input tokens, screening 200,000 resumes/month (avg 2K tokens) costs $30 in input. Output is short (structured score + 3-sentence summary), keeping total cost under $50/month for most hiring funnels. Claude Sonnet is preferable when the job description is unusually complex or when misclassification rates from the cheaper model prove too high in A/B testing.

Output schema should include: match score (0–100), top 3 matching qualifications, top 2 gaps, a single recommendation string. Never output a binary yes/no — recruiters need to calibrate their own threshold.

Common pitfalls

• Not including explicit instructions to ignore candidate name, address, and graduation year — models may inadvertently encode demographic proxies in their scoring
• Evaluating model output only against hiring decisions, not against recruiter ratings — the model should optimize for recruiter workload reduction, not final offer rate
• Using a single score across all roles — a software engineer role and a sales role have entirely different fit dimensions; use role-specific prompt templates
• Failing to audit output distributions monthly for disparate impact across protected groups; many jurisdictions (NYC, EU) require this for algorithmic hiring tools